See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/344982260

Development of a Robotic System for Automated Decaking of 3D-Printed Parts

Conference Paper · May 2020

DOI: 10.1109/ICRA40945.2020.9197110

CITATIONS READS
2 44

6 authors, including:

Huy Nguyen
Nanyang Technological University
7 PUBLICATIONS   25 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Large Scale Selective Laser Melting View project

Industrial Robot Calibration View project

All content following this page was uploaded by Quang Cuong Pham on 03 November 2020.

The user has requested enhancement of the downloaded file.

2020 IEEE International Conference on Robotics and Automation (ICRA)
31 May - 31 August, 2020. Paris, France

Development of a Robotic System for

Automated Decaking of 3D-Printed Parts
Huy Nguyen1,2* , Nicholas Adrian1,2 , Joyce Lim Xin Yan1,2 , Jonathan M. Salfity1,3 , William Allen1,3 and Quang-Cuong Pham1,2
1
HP-NTU Digital Manufacturing Corporate Lab, Nanyang Technological University, Singapore.
2
School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore.
3
HP Labs, HP Inc.
*
Corresponding author. Email: huy.nguyendinh09@gmail.com

Abstract—With the rapid rise of 3D-printing as a competitive

mass manufacturing method, manual “decaking” – i.e. removing
the residual powder that sticks to a 3D-printed part – has
become a significant bottleneck. Here, we introduce, for the
first time to our knowledge, a robotic system for automated
decaking of 3D-printed parts. Combining Deep Learning for
3D perception, smart mechanical design, motion planning, and
force control for industrial robots, we developed a system
that can automatically decake parts in a fast and efficient
way. Through a series of decaking experiments performed on
parts printed by a Multi Jet Fusion printer, we demonstrated
the feasibility of robotic decaking for 3D-printing-based mass
manufacturing.
Index Terms—deep learning, manipulation, system design,
3D-printing, decaking

I. I NTRODUCTION
With the rapid rise of 3D-printing as a competitive mass
manufacturing method, the automated processing of 3D-
printed parts has become a critical and urgent need. Post-
processing includes handling, polishing, painting, assembling
the parts that have been previously 3D-printed. In 3D-printing
processes that involve a powder bed, such as binder jet [1],
[2], Selective Laser Sintering, Selective Laser Melting [3] or
HP Multi Jet Fusion (MJF) [4], a crucial post-processing step
is decaking. After printing, some non-bound/non-melted/non-
Figure 1. Top: An operator removing powder from a 3D-printed part
sintered/non-fused powder usually sticks to the 3D-printed (manual decaking). Bottom: Our proposed robotic system for automated
part, forming together a “cake”. Decaking consists of remov- decaking. A video of the actual robotic decaking process can be found with
ing that residual powder from the part, which can then be fed the accompanying material of this paper.
to downstream post-processing stages.
1) Pick a caked part from the origin container with a
Decaking is currently done mostly by hand: human op-
suction cup;
erators remove the residual powder manually with brushes,
2) Clean the underside of the caked part by rubbing it on
as shown in Figure 1, top. Recent 3D-printing technologies,
a brush;
such as HP MJF, enable printing hundreds, or even thou-
3) Flip the caked part;
sands, of parts in a single batch, making manual decaking
4) Clean the other side of the caked part;
a major bottleneck in 3D-printing-based mass manufacturing
5) Place the cleaned part into the destination container.
processes. There are some automated approaches to decaking,
using for instance tumblers [5], but such approaches are Note that those steps are very general and can be applied to
unpractical if the powder is moderately sticky or if the parts other 3D-printing processes and to most parts that are nearly
are fragile. flat.
Here, we introduce, for the first time to our knowledge, a Step 1 is essentially a bin-picking task, which has recently
robotic system for automated decaking of 3D-printed parts, drawn significant attention from the community. However,
see Figure 1, bottom. Specifically, our system performs the the task at hand presents unique difficulties: (i) the caked
following steps: parts contain unpredictable amount of residual powder, over-
978-1-7281-7395-5/20/$31.00 ©2020 IEEE 8202
lap each other, and are mostly white-colored (the color of the
powder), making part detection and localization particularly
challenging; (ii) the residual powder and the parts have dif-
ferent physical properties, making caked part manipulation,
especially by a position-controlled industrial robot 1 , diffi-
cult. We address these challenges by leveraging respectively
(i) recent advances in Deep Learning for 2D/3D vision; and
(ii) smart mechanical design and force control.
Steps 2, 3 and 4 are specific to the decaking problem at
hand. A particular challenge is the control of the contacts be-
tween the industrial robot, the parts, and the brushing system.
We address this challenge by emphasizing smart mechanical
design whenever possible (cleaning station, flipping station),
and by using force control [6] to perform compliant actions
when necessary. Our pipeline can be applied to all parts with
nearly flat shapes, such as shoe insoles2 , which, to date, are
popular in 3D-printing-based mass manufacturing processes.
The remainder of the paper is organized as follows. In
Section II, we review related work in robotic cleaning, bin-
picking, 3D perception, and force control. In Section III,
we present our hardware system design. In Section IV,
we describe our software stack. In Section V, we evalu-
ate the system performance in number of actual decaking
experiments, and demonstrate the feasibility of automated
decaking. Finally, in Section VI, we conclude and sketch
some directions for future research.
II. R ELATED WORK
Recent years have seen a growing interest in robotic sur-
face/part cleaning. Advances in perception, control, planning
and learning have been applied to the problem [7]–[10]. Most
research used robot manipulators to clean particles, powder
or erase marker ink [7], [9]–[11], some considered mobile
robots to vacuum dust, or objects lying on the floor, such as
papers and clips [12], [13]. In fact, factors like variation in
type of dust, shape of object, material, and environmental
constraints require different approaches and techniques to
tackle the problem. In this work, we are interested in the
problem of cleaning 3D-printed parts which are randomly
placed in a bin, and each part is covered by an unknown
amount of residual or loose powder. Hence, we confine our
review of prior work as follows:
As mentioned previously, the first step of our task is
essentially a bin-picking task, which is one of the classical
problems in robotics and have been investigated by many
research groups, e.g. [14]–[16]. In these works, simplified
conditions are often exploited, e.g. parts with simple geo-
metric primitives, parts with holes to ease grasping action,
parts with discriminative features.
1 Most robots currently used in the industry are position-controlled, that is,
they achieve highly accurate control in position and velocity, at the expense
of poor, or no, control in force and torque. Yet, force or compliant control is
crucial for contact tasks, such as part manipulation and decaking. A number
of compliant robots have been developed in recent years but, compared to
existing industrial robots, they are still significantly more expensive, less
robust and more difficult to maintain.
2 The authors have been granted permission by Footwork Podiatry Labo-
ratory to use their shoe insole design and images in this paper.
8203
Since the Amazon Picking Challenge 2015, a variety of
approaches to a more general bin-picking problem had been
proposed [17]–[19]. In [20], authors discussed their obser-
vations and lessons drawn from a survey conducted among
participating teams in the challenge. Based on these lessons,
we identified 3D perception, which recognizes objects and
determining their 3D poses in a working space, as a key
component to build a robust bin-picking system. Histori-
cally, most approaches to conduct instance object recognition
are based on local descriptors around points of interest in
images [21]. While these features are invariant to rotation,
scale, and partial occlusions, they are not very robust to
illumination changes or partial deformations of the object.
Furthermore, since they depend on interest points and local
features, they only work well for objects with texture. Recent
advances in machine learning using convolutional neural
networks (CNNs) have rapidly improved object detection
and segmentation results [22]–[25]. Our detection module is
based on a Mask R-CNN model, which was originally open-
sourced by Matterport Inc., and implemented in TensorFlow
and Keras [23]. The Mask R-CNN model is an extension of
the Faster R-CNN [22] model that achieved rapid object clas-
sification and bounding-box regression. Mask R-CNN adopts
its two-stage procedure from Faster R-CNN. The first stage,
called Regional Proposal Network (RPN), proposes candidate
regions in which there might be an object in the input image.
The second stage, in parallel, performs classification, refines
the bounding-box and generates a segmentation mask of the
object on each proposal region. Both stages are connected to
a backbone structure.
Another important conclusion is that system integration
and development are fundamental challenges to build task-
specific, robust integrated systems. It is desirable to study
integrated solutions which include all component technolo-
gies, such as 3D object pose estimation, control, motion
planning, grasping, etc. Besides, it is also highlighted that
good mechanical designs can sidestep challenging problems
in control, motion planning and object manipulation. In line
with this idea, our approach combines 3D perception using
Deep Learning, smart mechanical design, motion planning,
and force control for industrial robots, to develop a system
that automatically decakes parts in a fast and efficient way.
In addition, another challenging problem that is unique
to our decaking task is force control. To achieve compliant
motion control, we take a position-controlled manipulator
as a baseline system and make necessary modifications by
following the pipeline in [26], [27] (thanks to its robustness).
This enables safe and sufficient cleaning operations without
the risk of damaging the parts. For more specific surveys on
force control, interested readers can refer to [28], [29].
III. H ARDWARE SYSTEM OVERVIEW
We design our system while keeping in mind a scalable
solution that could eventually tackle the problem of post-
processing 3D-printed parts in a real scenario, at an advanta-
geous cost. The robotic platform used in this work is, there-
fore, characterized by cost-efficient, off-the-shelf components
 Figure 2. Our hardware system design

combined with a classical position-control industrial manip-

ulator. In particular, the main components of our platform
Figure 3. Cleaning station, comprising of a fan, a brush rack, and a vacuum
are outlet.
• 1 Denso VS060: Six-axis industrial manipulator;
• 1 ATI Gamma Force-Torque (F/T) sensor;
D. Flipping station
• 1 Ensenso 3D camera N35-802-16-BL;
• 1 suction system powered by a Karcher NT 70/2 vacuum
machine;
• 1 cleaning station;
• 1 flipping station.
All control and computations are done on a computer with
an Intel Xeon E5-2630v3, 64 GB RAM.
A. Suction system
For grasping the parts, we decided to used a suction cup.
This choice was motivated by the versatility and performance
of suction-based system for bin-picking [20], and by fact that
most recovered powders are recyclable. Our suction system is
designed to generate both high vacuum and high air flow rate.
Figure 4. Flipping station.
This is to provide sufficient force to lift parts and to maintain
a firm hold of parts during brushing. Air flow is guided
To change the part orientation, we implement a passive
from a suction cup on the tip of the end effector through
flipping station (i.e. no actuator is required to perform
a flexible hose into the vacuum machine. The vacuum itself
flipping). As illustrated in Figure 4, the part is dropped from
is generated by a 2400W vacuum cleaner. For binary on/off
the top of the station and moves along the guiding sliders.
control, we customize a 12V control output from the DENSO
Upon reaching the bottom of the slider, the part is flipped
RC8 controller.
and is ready to be picked up by the robot.
B. Camera This design is simple but works well with relatively
Robust perception is a one of the key components of flat parts such as shoe insoles, flat boxes, etc. The next
the system. We manually optimized the selection of camera improvement of the station will accommodate general part
and camera location to maximize the view angles; avoid geometry and provide more reorientation options.
occlusions from robot arm, end-effector and cleaning, pick- IV. S OFTWARE SYSTEM OVERVIEW
ing, dropping stations; avoid collisions with objects and
environment; and safety of the device. Our software system is composed of a state machine and
For the sake of robustness, our system employs Ensenso a series of modules. The communcation between different
3D camera N35-802-16-BL and combines the use of both 2D hardware and software components is based on the Robot
grayscale image and depth information. Operating System (ROS) [30].

C. Cleaning station A. State machine

To clean the parts, a brush rack is installed at the bottom The state machine is responsible for selecting the appro-
of the cleaning station. We also integrate a dust management priate module to execute at each point in time, see Figure 5.
system to collect excess powder removed during brushing. The solid arrows show the direction to next states depending
Throughout the cleaning process, a blowing fan is automati- on the result of current states. The state machine has access
cally activated to direct the powder stream frontward into the to all essential information of the system, including types,
vacuum suction hose, see Figure 3. poses, geometries and cleanliness, etc. of all objects detected
8204

Part
segmentation
Initialization Picking
and
localization
Figure 5. The system is composed of a state machine and a series of modules to perform perception and different types of action.
Mask R-CNN
Figure 6. Example result of the object detection module based on Mask R-CNN. The estimated bounding boxes and part segmentations are depicted in
different colors and labelled with the identification proposal and the confidence. We reject the detection with confidence lower than 95%
in the scene. Each module can query this information to
realize its behavior. As a result, this design is general and
can be adapted to many more types of 3D-printed parts.
For example, the system can choose and execute a suitable
cleaning strategy based on the type of part, its pose, and its
prioritized list of actions; and decide whether to reattempt
with a different strategy or skip that part based on the
outcome. In our implementation, the modules were fine-
tuned for each type of object based on previous experiments
to obtain a high success rate. It is desirable to have a
systematic approach to optimize the algorithm which decides
the best cleaning strategy to use for each type of part (see
Section V-C).
The state machine is also responsible for selecting the most
feasible part to be decaked next. Since the parts are often
cluttered or stacked, we generate the cleaning sequence by
prioritizing parts with large top surfaces and are on top of
the clutter. This is to perform suction easier and to de-stack
parts in cluttered configurations.
B. Perception
The perception task is to identify and localize visible
objects in the working space. However, this is a challenging
problem, due to heavy occlusions (i.e. self-occlusion and
occlusion by powder) and the poor contrast between the
freshly printed objects and their background. Our approach
is to separate the perception into two stages: object detection,
segmentation and 3D pose estimation. At the first stage, we
utilize a state-of-the-art deep-learning network to perform in-
stance detection and segmentation. The second stage extracts
the 3D points of each object using the segmentation mask
and estimate the object pose. This pipeline allow us to exploit
the strength of deep-learning in processing 2D images and
depth information in estimating object pose. We believe such
Y
Y
Both sides  Caked parts
Cleaning Stowing
cleaned? remaining?
N N
Flipping End
Pointcloud extraction
combination of depth and RGB to be essential for a robust
perception solution.
First, a deep neural network based on Mask R-CNN
classifies the objects in the RGB image and performs instance
segmentation, which provides pixel-wise object classification.
The developers had provided a Mask R-CNN model that was
pre-trained on large image classification datasets (Microsoft
COCO dataset [31]). Hence, we applied transfer learning
on the pre-trained model to save training time, by training
the network to be able to classify a new object class. In
our experiment, this network requires from 75-100 labelled
images of real occurrences of typical 3D-printed parts. The
result was at very high detection rate of all parts presenting
in the bin, as shown in Figure 6. On average, our network
has a recall of 0.967 and a precision of 0.975.
Second, pose estimation of the parts is done by estimating
the bounding boxes and computing the centroids of the
segmented pointclouds. The pointcloud of each object is
refined (i.e. statistical outlier removal, normal smoothing,
etc.) and used to verify if the object can be picked by suction
(i.e. exposed surfaces must be larger than suction cup area).
The 3D information of the bin is also used later during
the motion planning for collision avoidance. We included
heuristics to increase the robustness of the estimations by
specifying physical constraints inside the bin e.g. parts must
lie within the bin’s walls and cannot be floating within the
bin.
We also note that the Ensenso camera cannot capture both
2D and 3D images simultaneously as the high-intensity IR
light tends to blind the 2D camera. This module is, therefore,
also in charge of turning on and off the high-intensity IR
strobe to prevent it from affecting 2D images.
8205
C. Motion primitives
Some modules, such as “Picking” and “Cleaning”, are
themselves composed of a series of motion primitives.
Picking (suction-down) motion primitives: These primi-
tives are useful for picking parts with exposed nearly flat
surfaces. We limit our implementation to top-down direc-
tions. The process is as follows:
1) The robot moves the suction cup above the centroid of
the part and lowers the cup. Compliant force control is
enabled and tells the robot when to stop the moving-
down motion.
2) Check whether the height at which the suction cup was
stopped matches the expected height (from the point
cloud with some tolerances).
3) The suction cup lifts, and the system constantly checks
the force torque sensor to determine whether there is
any collision (which often results large drag forces).
Using suction is typically superior for parts with nearly flat
and wide surfaces. However, even when the suction cup could
not form a perfect seal with porous parts, high air flow often
maintains sufficient force on the part to lift it. Moreover, the
suction cup is versatile and can be replaced to fit many types
of object.
Cleaning motion primitives: These primitives remove
residual powder and debris from the printed parts. Our current
implementation of these primitives is limited to parts that are
nearly flat (e.g. shoe insoles).
First, the robot positions the part above the brush rack in
the cleaning station. Then, the compliant force control mode
is enabled to move the robot until contact with the brushes.
Next, the cleaning trajectories are performed. To maintain
constant contact between the brushes and the part, we apply
a hybrid position/force control scheme: force is regulated in
the direction normal to the brushes’ surfaces, while position
is regulated in the tangential directions. The force thresholds
are determined through a number of trial-error experiments.
The cleaning trajectories are planned following two patterns:
spiral and rectircle (see Figure 7). While the spiral motion
is well-suited for cleanning nearly flat surfaces, the rectircle
motion aids with removing powder in concave areas.
D. Motion planning and control
All collision-free movements are planned online using
the Bi-directional Rapidly-Exploring Random Trees (RRTs)
algorithm [32] available in OpenRAVE [33].
V. S YSTEM PERFORMANCE AND DISCUSSION
A. Experimental setup
Parts considered in the experiments were 3D-printed shoe
insoles. At the current stage of development, we assumed
all parts to be unpacked from printer powderbed and placed
randomly in container bin. The task of the system was to
remove powder from the parts, which naturally consisted of
several key steps as follows:
1) Initial localization of parts in container bin,
2) Picking a part out of the container bin,
8206
Figure 7. We use a combination of spiral and rectircle paths for cleaning
motions. An example of spiral paths is shown in red. The yellow dot denotes
the centroid of the parts at beginning of the motion. We modify the spiral
paths so that they continue to circle around this yellow dot after reaching a
maximum radius. An example of rectircle path is shown in blue. Rectircle
path’s parameters include width, height and its direction in XY plan.
3) Removing residual powder from the underside of the
part,
4) Flipping the part,
5) Re-localizing parts in the container bin and in the
flipping collection area,
6) Picking flipped part,
7) Removing residual powder from the other side of the
part,
8) Placing the part into the collection bin and continuing
with any remaining parts (Step 2-8).
B. Performance
The sytem performed the cleaning task on 10 freshly 3D-
printed shoe insoles. Two aspects of the proposed system
were experimentally investigated. First, we evaluated the
cleaning quality by weighing the parts before and after
cleaning. Second, we reported the actual running time of the
proposed system in a realistic setting. As the baseline for
comparison, we also considered the same task performed by
skilled human operators. The results of this experiment are
summarized in Table I.
As expected, the performance of the system yielded a
promising result as a first prototype. Two test runs were con-
ducted and successfully executed by the proposed system. We
noted that a perfectly cleaned printed shoe insoles weighed
22.4 ± 2g on average. The proposed system could remove
powder from a dirty part on an average of 50.1 ± 10.9s and
parts weighed 37.6 ± 6.4g on average after cleaning, which
meant 42.0% ± 24.4% excess powder has been removed
from the parts. In comparison, a skilled human operator
could clean 72.0% ± 12.1% of excess powder (parts weighed
29.8±3.2g after cleaning in 41.2±1.9s on average). Although
our system performance regarding the cleaning quality was
1.7 times less than that of a human operator, this was a
potential outcome as our robot system is supposed to work
significantly longer. This however raised questions how task
efficiency could be further improved.
C. Discussion Table I
P ERFORMANCE OF THE PROPOSED SYSTEM , AS COMPARED TO THAT OF
We observed that our system performed actual brushing SKILLED HUMAN OPERATORS .
actions in only 40% of the execution time. In comparison,
human spent more than 95% execution time on brushing. Avg per part Our system Human (no Human (20s
time-limit) brushing)
This was attributed to human superior skills in performing
Mass Before (g) 48.6 ± 10.9 48.8 ± 7.8 48.9 ± 8.0
sensing and dexterous manipulations. However, when we Mass After (g) 37.6 ± 6.4 29.8 ± 3.2 34.2 ± 4.5
limited brushing time to 20s, cleaning quality was reduced as Cycle Time (s) 50.1 ± 2.1 41.2 ± 1.9 21.2 ± 1.1
parts weighed 34.2±4.5g on average after cleaning (removed Brushing Time (s) 20 40 20
55.5%±17.9% excess powder). This suggested that one could
improve cleaning quality by prolonging the brushing duration
and upgrading the cleaning station (e.g. spinning brushes to
speed up cleaning action; compliant brushes to clean concave
areas, etc.).
We also noted that humans provided more consistent re-
sults of cleaned parts (i.e. parts cleaned by human operators
had weight variation much smaller than that of parts cleaned
by our system). This was due to the fact that human operators
Figure 8. Average time-line representation of all actions used for the
actively adjusted their cleaning motions based on vision feed- cleaning task.
back. Hence, incorporating a cleanliness evaluation module
into the current system would improve the cleaning quality goal is not to pick, but to change the configuration of an
performance. For this purpose, another 3D camera could be object that cannot be picked by the other primitives. It is
used to observe the object after each cleaning section. The called in particular when the exposed face of an object is
module would retrieve a list of all dirty locations on object too narrow. In this scenario it may be possible to topple the
surfaces. This information would allow the system to plan object into a new configuration that will allow the object to
better cleaning motions and could also be utilized to validate be picked by suction. Finally, air flow can be measured by
cleaning results. using a pitot tube inside the suction cup/hose. This is used to
Moreover, though this experiment featured a batch of detect whether a part was successfully grasped or lost during
nearly the same 3D-printed shoe insoles, our ultimate goal arm motions.
would be to extend this robotic system to a various types
of 3D-printed parts. For example, in many cases, a batch of
3D-printed parts may contain a number of different parts.
The cleaning routines, hence, would have to be altered to
cater to the different geometries of each parts. To achieve VI. C ONCLUSION
that, the state machine can be encoded with specific cleaning
instructions for each type of part. Much in the same way In this paper, we have presented an automated robotic
human operators change cleaning routines for different parts, system for decaking 3D-printed parts. We have combined
our robotic system may recognize different parts, look up a Deep Learning for 3D perception, smart mechanical design,
prior defined handling instructions, perform the cleanliness motion planning and force control for industrial robots to
evaluation and execute accordingly. develop a system that can perform decaking task in a fast and
To investigate the computational cost of our approach, efficient manner. Through a series of experiments performed
Figure 8 also provides information on the average times of on parts printed by a HP Multi Jet Fusion printer, we
each action used when performing the task. We noted that our demonstrated the feasibility of automatizing such tasks. The
robot ran at 50% max speed and all motions were planned achievements of this first prototype are promising and lay
online. Hence, the sytem performance could be further en- the groundwork for multiple possible extensions: to other
hanced by optimizing these modules (e.g. off-line planning types of parts, to other powder-based 3D-printing processes
for tight-space motions, working space optimization, etc.). (SLM, SLS, binder jet, etc.) In future work, we will focus on
Moreover, our perception module was running on a CPU, validating the current system further and expanding it to more
implementations of better computing hardware would thus general scenarios e.g. incorporating a cleanliness evaluation
improve the perception speed. module, optimizing task efficiency and adapting the system
Furthermore, we are also working forward improving our to work on more general parts.
end-effector design. The current system is only capable of
performing suction-down motion primitives. Though it can ACKNOWLEDGMENT
accommodate for a very large tilting angle, our system does
not have enough maneuverability to perform side picking. A This research was conducted in collaboration with HP Inc.
possible approach would be to design a more generic suction and partially supported by the Singapore Government through
gripper to perform the task. Another solution would be to add the Industry Alignment Fund -Industry Collaboration Projects
another assisting motion primitive, which is Toppling whose Grant.
8207
 R EFERENCES
[1] E. M. Sachs, J. S. Haggerty, M. J. Cima, and P. A. Williams, “Three-
dimensional printing techniques,” Apr. 20 1993, US Patent 5,204,055.
[2] Y. Bai and C. B. Williams, “An exploration of binder jetting of copper,”
Rapid Prototyping Journal, vol. 21, no. 2, pp. 177–185, 2015.
[3] E. O. Olakanmi, R. Cochrane, and K. Dalgarno, “A review on selective
laser sintering/melting (SLS/SLM) of aluminium alloy powders: Pro-
cessing, microstructure, and properties,” Progress in Materials Science,
vol. 74, pp. 401–477, 2015.
[4] hp.com. (2019) HP multi jet fusion technology. [Online].
Available: https://www8.hp.com/sg/en/printers/3d-printers/products/
multi-jet-technology.html
[5] A. Ju, A. Fitzhugh, J. Jun, and M. Baker, “Improving aesthetics through
post-processing for 3D printed parts,” Electronic Imaging, vol. 2019,
no. 6, pp. 480–1, 2019.
[6] G. Zeng and A. Hemami, “An overview of robot force control,”
Robotica, vol. 15, no. 5, pp. 473–482, 1997.
[7] F. Sato, T. Nishii, J. Takahashi, Y. Yoshida, M. Mitsuhashi, and
D. Nenchev, “Experimental evaluation of a trajectory/force tracking
controller for a humanoid robot cleaning a vertical surface,” in 2011
IEEE/RSJ international conference on intelligent robots and systems,
2011, pp. 3179–3184.
[8] F. Nagata, T. Hase, Z. Haga, M. Omoto, and K. Watanabe,
“CAD/CAM-based position/force controller for a mold polishing
robot,” Mechatronics, vol. 17, no. 4-5, pp. 207–216, 2007.
[9] J. Hess, G. D. Tipaldi, and W. Burgard, “Null space optimization for
effective coverage of 3d surfaces using redundant manipulators,” in
IEEE/RSJ International Conference on Intelligent Robots and Systems,
2012, pp. 1923–1928.
[10] C. Eppner, J. Sturm, M. Bennewitz, C. Stachniss, and W. Burgard,
“Imitation learning with generalized task descriptions,” in International
Conference on Robotics and Automation. IEEE, 2009, pp. 3968–3974.
[11] J. Hess, M. Beinhofer, and W. Burgard, “A probabilistic approach to
high-confidence cleaning guarantees for low-cost cleaning robots,” in
International conference on robotics and automation (ICRA). IEEE,
2014, pp. 5600–5605.
[12] G. Lawitzky, “A navigation system for cleaning robots,” Autonomous
Robots, vol. 9, no. 3, pp. 255–260, 2000.
[13] J. L. Jones, “Robots at the tipping point: the road to irobot roomba,”
IEEE Robotics & Automation Magazine, vol. 13, no. 1, pp. 76–78,
2006.
[14] B. Drost, M. Ulrich, N. Navab, and S. Ilic, “Model globally, match
locally: Efficient and robust 3d object recognition,” in Computer
society conference on computer vision and pattern recognition. IEEE,
2010, pp. 998–1005.
[15] D. Buchholz, D. Kubus, I. Weidauer, A. Scholz, and F. M. Wahl,
“Combining visual and inertial features for efficient grasping and bin-
picking,” in International Conference on Robotics and Automation
(ICRA). IEEE, 2014, pp. 875–882.
[16] A. Pretto, S. Tonello, and E. Menegatti, “Flexible 3D localization
of planar objects for industrial bin-picking with monocamera vision
system,” in International Conference on Automation Science and
Engineering (CASE). IEEE, 2013, pp. 168–175.
8208
View publication stats
[17] C. Eppner, S. Höfer, R. Jonschkowski, R. Martı́n-Martı́n, A. Sieverling,
V. Wall, and O. Brock, “Lessons from the amazon picking challenge:
Four aspects of building robotic systems.” in Robotics: Science and
Systems, 2016.
[18] C. Hernandez, M. Bharatheesha et al., “Team delft’s robot winner of
the amazon picking challenge 2016,” in Robot World Cup. Springer,
2016, pp. 613–624.
[19] K.-T. Yu, N. Fazeli, N. Chavan-Dafle, O. Taylor, E. Donlon, G. D.
Lankenau, and A. Rodriguez, “A summary of team MIT’s approach to
the amazon picking challenge 2015,” arXiv preprint arXiv:1604.03639,
2016.
[20] N. Correll, K. E. Bekris, D. Berenson, O. Brock, A. Causo, K. Hauser,
K. Okada, A. Rodriguez, J. M. Romano, and P. R. Wurman, “Lessons
from the amazon picking challenge,” arXiv preprint arXiv:1601.05484,
vol. 3, 2016.
[21] D. Lowe, “Object recognition from local scale-invariant features.” in
International Conference on Computer Vision, vol. 99, no. 2. IEEE,
1999, pp. 1150–1157.
[22] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards real-
time object detection with region proposal networks,” in Advances in
neural information processing systems, 2015, pp. 91–99.
[23] K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask R-CNN,” in
Proceedings of the IEEE international conference on computer vision,
2017, pp. 2961–2969.
[24] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu,
and A. C. Berg, “SSD: Single shot multibox detector,” in European
conference on computer vision. Springer, 2016, pp. 21–37.
[25] J. Redmon and A. Farhadi, “YOLO9000: better, faster, stronger,” in
Proceedings of the IEEE conference on computer vision and pattern
recognition, 2017, pp. 7263–7271.
[26] F. Suárez-Ruiz, X. Zhou, and Q.-C. Pham, “Can robots assemble an
IKEA chair?” Science Robotics, vol. 3, no. 17, p. eaat6385, 2018.
[27] C. Ott, R. Mukherjee, and Y. Nakamura, “Unified impedance and
admittance control,” in International Conference on Robotics and
Automation. IEEE, 2010, pp. 554–561.
[28] T. Lefebvre, J. Xiao, H. Bruyninckx, and G. De Gersem, “Active
compliant motion: a survey,” Advanced Robotics, vol. 19, no. 5, pp.
479–499, 2005.
[29] A. Calanca, R. Muradore, and P. Fiorini, “A review of algorithms for
compliant control of stiff and fixed-compliance robots,” IEEE/ASME
Transactions on Mechatronics, vol. 21, no. 2, pp. 613–624, 2015.
[30] M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs,
R. Wheeler, and A. Y. Ng, “ROS: an open-source robot operating
system,” in ICRA workshop on open source software, vol. 3, no. 3.2.
Kobe, Japan, 2009, p. 5.
[31] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan,
P. Dollár, and C. L. Zitnick, “Microsoft COCO: Common objects in
context,” in European conference on computer vision. Springer, 2014,
pp. 740–755.
[32] J. J. Kuffner Jr and S. M. LaValle, “RRT-connect: an efficient ap-
proach to single-query path planning,” in International Conference on
Robotics and Automation, vol. 2. IEEE, 2000.
[33] R. Diankov, “Automated construction of robotic manipulation pro-
grams,” Ph.D. dissertation, Carnegie Mellon University, 2010.